Right Ventricle Segmentation in Short-Axis MRI Using a Shape Constrained Dense Connected U-Net

Segmentation of right ventricle (RV) in short-axis MRI is an essential step for evaluating the structure and function of RV. In this paper, a shape constrained deep learning network based on dense connectivity and dilated convolutions is proposed, which aims to strengthen feature propagation and have more diversified features through dense connections and skip connections. In the meantime, dilated convolution is used to expand the receptive fields and enhance the connectivity of segmentation results. The shape constraint of RV is introduced into the loss function to improve the prediction accuracy. Transfer learning is employed to strengthen the generalization of the network to the RV shape constraint. Finally, post-processing is performed by analysing boundary curvature of segmentation results and shape correlation of endocardium and epicardium of RV. Our network is validated on the MICCAI2012 public dataset, and the evaluation results show that our network outperforms the state-of-the-art methods in several evaluation metrics.